Enhancing protein aggregation prediction: a unified analysis leveraging graph convolutional networks and active...

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Structural Biology

Background:

• Protein aggregation (PA) is implicated in neurodegenerative diseases like Alzheimer's and Parkinson's.
• Understanding PA requires insights into aggregation-prone regions (APRs) and structural interactions.
• Computational methods, especially machine learning, offer efficient alternatives to experimental PA studies.

Purpose of the Study:

• To develop a Graph Convolutional Network (GCN) model for accurate protein aggregation (PA) score prediction.
• To leverage expanded datasets from Protein Data Bank (PDB) and AlphaFold2.0 for improved model training.
• To evaluate an active learning strategy for efficient identification of proteins with high PA propensity.

Main Methods:

• Constructed a GCN model utilizing an enhanced dataset derived from PDB and AlphaFold2.0.
• Calculated PA propensity using AGGRESCAN3D 2.0 and refined PDB data by separating multi-polypeptide chains.
• Incorporated 22,774 Homo sapiens sequences from AlphaFold2.0 after sequence similarity comparison.

Main Results:

• The trained GCN model achieved a high coefficient of determination (R ²) of 0.9849 and a low mean absolute error (MAE) of 0.0381 for PA prediction.
• The active learning approach demonstrated superior performance with an MAE of 0.0291 in expected improvement.
• Active learning identified 99% of target proteins by exploring only 29% of the search space.

Conclusions:

• The developed GCN model shows significant promise for predicting protein aggregation susceptibility.
• The active learning strategy enhances the efficiency of identifying proteins prone to aggregation.
• This work advances computational tools for PA prediction, with potential applications in disease diagnosis and therapy.